Pith. sign in

REVIEW 3 major objections 5 minor 37 references

Beyond Sufficiency: Time Series Explanation with Counterfactual Necessity

T0 review · 3 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that a faithful time-series explanation must be counterfactually necessary, not merely sufficient, and presents TimePNS, a two-stage framework that derives necessity signals from latent causal interventions and uses them t

desk verdict TimePNS is a genuinely new combination—latent counterfactual necessity as a refinement signal for time-series masks—and the empirical story is consistent, but the causal core rests on an unvalidated local linearization and unstated identifiability conditions. read the letter →

arxiv 2607.21573 v1 pith:6NS35DHT submitted 2026-07-23 cs.LG cs.AI

Hongnan Ma , Yiwei Shi , Mengyue Yang , Weiru Liu This is my paper · ORCID
classification cs.LGcs.AI
keywords time-seriesexplanationcounterfactualnecessityprobabilityofandsufficiencycausalrepresentationlearninglatentinterventionsmaskstemporalsaliency
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Most perturbation-based time-series explainers select subsequences that are sufficient to keep the model's prediction, but sufficiency alone rewards correlated, non-essential subsequences. The paper argues that a faithful explanation must also be necessary: if the selected subsequence is removed, the prediction should change. To operationalize this, TimePNS first learns an identifiable causal generative process over latent temporal factors, alongside a sufficiency-oriented mask (Stage I). It then performs latent counterfactual interventions — setting one latent factor to a reference value and propagating the perturbation through learned lagged and instantaneous dependencies — to compute a necessity score for every factor, and trains a temporal gate to align the mask with those scores (Stage II). On synthetic and real-world benchmarks (ECG, epilepsy, word recognition), the paper reports that TimePNS identifies decision-critical subsequences more accurately and improves the sufficiency–necessity trade-off relative to sufficiency-oriented baselines.

What carries the argument

The central object is the latent counterfactual propagation rule (Eq. 6), which linearizes the residual-preservation constraint around the factual latent sequence. For an intervention on latent factor (τ,k), all other factors' recovered residuals are held at their factual values, and the first-order Jacobians of the inverse transition mechanisms (J_lag, J_inst) propagate the perturbation Δz through both lagged and instantaneous dependencies under a topological ordering. The resulting counterfactual latent sequence is decoded and scored by the black-box model, yielding the necessity score Δp. This propagation rule is what lets TimePNS estimate counterfactual necessity in latent space without

What would settle it

A concrete disconfirming observation: if TimePNS is working, then on a dataset where a purely predictive but non-causal channel (e.g., a channel that is a noisy copy of the true causal channel) is added to the input, the necessity scores must consistently assign near-zero necessity to that channel while sufficiency scores remain high. A direct experiment: train a classifier on data with a redundant channel, run TimePNS, and ablate only the redundant channel — the prediction must remain unchanged; if removing the redundant channel alone changes the prediction, the necessity signal is not captur

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that counterfactual necessity can be computed from a learned latent causal model and used as a supervision signal that refines sufficiency-based explanations. The key result is that necessity, measured as the drop in factual-class confidence after a latent intervention (Δp in Eq. 7), systematically downweights predictive-but-redundant temporal factors — such as the reciprocal ST-depression correlated with ST-elevation in ECGs — that sufficiency masks keep. In experiments, the paper reports that TimePNS's masks outperform sufficiency baselines on ground-truth saliency metrics (AUPRC, AUP) and on occlusion-based sufficiency/necessity tests, an

Load-bearing premise

The necessity scores rest on two linked assumptions: that the latent causal model learned in Stage I is identifiable from observational data (so that interventions in the learned latent space correspond to true counterfactuals of the data process), and that the first-order linearization of the residual-preservation constraint (Eq. 6) is accurate for the chosen reference intervention; if either fails, Δp does not measure genuine counterfactual necessity.

Editorial extensions

If this is right

  • If TimePNS's necessity signal works as claimed, sufficiency-only explainers systematically over-report correlated but non-essential subsequences as important; necessity refinement should reduce false positives in explanation masks.
  • The learned latent inverse mechanisms provide an intervention space that is causal-valid where raw input-space replacement is not, so necessity scoring can be applied to any frozen black-box time-series classifier without retraining it.
  • TimePNS's disagreement-set ablation implies that the positions it uniquely selects—those that sufficiency baselines drop—are the ones whose removal most changes the prediction; sufficiency masks are therefore partially redundant.
  • The case-study result — a sufficient mask whose removal causes almost no confidence drop (0.001) versus 0.514/0.740 for TimeX++/TimePNS — implies that sufficiency and necessity diverge widely on real data, so necessity evaluation should be added to standard occlusion-based protocols.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A testable consequence the paper does not run: if TimePNS's necessity masks are truly identifying causal features rather than predictive ones, they should be more stable under distribution shift (e.g., covariate shift in the non-essential channels). The introduction suggests this, but no shift experiment is reported; this is our inference, not the paper's claim.
  • The linearized propagation (Eq. 6) is only a first-order approximation and assumes J(0)_{i,i} ≠ 0 plus small intervention magnitude. A stress test comparing these scores against exact SCM counterfactuals (computed via full forward simulation) or against input-space occlusion would clarify whether the linearization is the limiting factor.
  • The necessity signal's validity depends on identifiability of the latent causal process from observational data. If the reference intervention z_ref is not faithful under the learned model, the scores Δp may measure 'what the learned model would do' rather than true necessity; the paper does not state identifiability conditions, so an identifiability audit is an open direction.
  • The manuscript defers several derivations (e.g., the linearized propagation derivation and full objective definitions) to 'Appendix??' placeholders; a complete version would need those details for the necessity signal to be auditable.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes TimePNS, a two-stage framework for post-hoc explanation of black-box time-series classifiers. Stage I jointly learns a latent causal model of the observed time series (with an inverse-mechanism formulation of lagged and instantaneous dependencies) and a sufficiency-oriented explanation mask that preserves the black-box prediction. Stage II computes per-factor counterfactual necessity scores by intervening on a single latent factor, propagating the intervention through the learned latent dynamics via a first-order linearization of the residual-preservation constraint, and decoding back to input space to measure the drop in the original prediction confidence. These scores supervise a temporal gate that refines the Stage I mask toward decision-critical subsequences. Experiments on three synthetic and three real-world datasets compare TimePNS with TimeX, TimeX++, Dynamask, CORTX, WinIT, and SGT, and report improved AUPRC on synthetic data as well as stronger occlusion-based sufficiency/necessity behavior on real data.

Significance. If the necessity signal is valid, the paper makes a useful conceptual contribution: moving time-series explanation from sufficiency alone to a Pearl-style necessity criterion, and doing so in a latent space to avoid off-manifold input perturbations. The paper has concrete strengths: a clear two-stage formulation, a well-motivated synthetic evaluation with ground-truth saliency, and ablations (Stage I vs. Stage II, disagreement-set analysis) that isolate the effect of the necessity-guided refinement. However, the central claim rests on the correctness of the latent counterfactual propagation in Sec. 5.3, and that step is currently under-validated. Identifiability of the latent process is also asserted rather than demonstrated. These are load-bearing issues: if the counterfactual necessity scores are systematically biased, the improved masks are not evidence that true counterfactual necessity has been captured.

major comments (3)
  1. [Sec. 5.3, Eqs. (5)-(7)] The sole necessity supervision Δp_{τ,k} is computed from a latent counterfactual trajectory obtained by first-order linearization of the residual-preservation constraint. This is a local approximation, yet the intervention sets z_{τ,k} to a reference value z̃_{τ,k} that may be far from the factual value, and the recursive propagation for s>τ compounds the error at every step. No experiment checks whether the resulting counterfactual latents satisfy the nonlinear residual constraint or whether decoded X_cf remains on the data manifold. Since the gate is trained against Δp_{τ,k}, a systematic propagation error directly biases the final mask. Please provide a validation of the linearization: for the same interventions, solve the nonlinear residual equations numerically (e.g., fixed-point iteration or gradient descent), report residual norms and the distance between linearized and exact coun
  2. [Sec. 5.1 and Sec. 6.1] The paper repeatedly states that Stage I learns an 'identifiable causal generative process' and cites Li et al. [2024], but it does not state the identifiability conditions or explain how the training objective in Eq. (8) satisfies them. If the latent factors are not identifiable up to the appropriate transformation, the 'factors' z_{τ,k} that are intervened upon are not well-defined units, and the necessity score Δp_{τ,k} cannot be interpreted as the effect of intervening on a real temporal factor. Please state the identifiability conditions from the cited work, and either prove or empirically demonstrate that the learned latent representation satisfies them in the experimental settings. If identifiability cannot be guaranteed, the causal language should be weakened accordingly.
  3. [Sec. 7.2, Table 1 and Fig. 2] The empirical claim of 'consistently improves sufficiency-necessity trade-offs' is stronger than the evidence shown. On synthetic data, TimePNS has the best AUPRC on all three datasets but is not the best AUR on any dataset and is not the best AUP on LowV-ER; the gains over TimeX/TimeX++ are small in several cases, and no statistical significance tests are reported. On real data, Fig. 2 shows curves without error bars or confidence intervals despite 5-fold cross-validation, and the text reports qualitative trends rather than numeric effect sizes. Please report the quantitative values, variability, and significance tests for the key real-world occlusion metrics, and soften the claim if the differences are not statistically robust.
minor comments (5)
  1. [Throughout] Several typos and placeholder artifacts remain: 'Dynamsk' for Dynamask, 'LowV-ER' vs. 'Low-VAR', 'Sbaseline ∖ TimePNSSTimePNS ∖ baseline' garbled in Fig. 3, and multiple 'Fig. ?? / Appendix??' references. These should be cleaned up.
  2. [Sec. 3] The mask matrix is defined as binary, but training uses stochastic masks and a straight-through estimator. The relationship between the relaxed probabilities π and the final reported binary mask should be clarified (e.g., thresholding at 0.5 or sampling at test time).
  3. [Sec. 5.4] The necessity score Δp_{τ,k} is clipped at zero but not normalized across factors or time steps; the soft target fPN is then aligned via centered cosine similarity in Sec. 6.2, which is shift-invariant and scale-invariant. This may make the alignment insensitive to the magnitude of necessity. A brief justification or an ablation on the target formulation would help.
  4. [Sec. 7.1] Hyperparameter choices for the black-box Transformer, the number of latent factors d, the lag L, the intervention reference value z̃_{τ,k}, and the evaluation time steps T_eval are deferred to appendices that are not included in this version. Since the method's behavior depends strongly on these choices (especially z̃ and T_eval), they should be stated in the main text or the appendices made available.
  5. [Sec. 7.2.2] The disagreement-set ablation uses top-25% selected positions and measures |Δp_GT|. It would be useful to state whether the baseline masks are thresholded in the same way as TimePNS masks, and to report the number of samples in each disagreement set to assess the reliability of the win rates.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the necessity signal is an external supervision target derived from latent counterfactual interventions, not from the final explanation mask.

full rationale

TimePNS's central necessity signal Δp_{τ,k} (Eq. 7) is computed by intervening on a learned latent factor, decoding the counterfactual trajectory, and querying the frozen black-box classifier. This quantity is defined independently of the final explanation mask M, so the mask refinement is not equivalent to the score it is trained against. The gate is supervised by Δp through L_align and the prediction-anchor loss, and the final mask is produced by injecting the gated latent into the explainer; evaluation then uses held-out occlusion tests on input positions. This is a standard train/evaluation loop rather than a self-definitional reduction. The latent causal model is adopted from an external citation (Li et al., 2024), not from the present authors, and no uniqueness theorem by the same authors is invoked. The only self-citations (Yang et al., 2023; Yu et al., 2025) appear in related-work context and are not load-bearing for the paper's claims. The first-order linearization in Eq. 6 is a legitimate correctness concern—it may bias the necessity scores—but it is not circularity, because the approximation is not constructed from the final mask or from the evaluation metric. No prediction in the paper reduces by construction to a fitted parameter or to the paper's own inputs.

Assumptions & free parameters 5 free parameters · 5 assumptions · 1 invented entities

The central claim rests on a latent causal process whose identifiability is inherited from a cited paper, a linearized counterfactual propagation step with no error control, and several hand-chosen hyperparameters and intervention references. The paper provides no independent verification that these latent entities correspond to the factors the black-box actually uses.

free parameters (5)
  • sparsity target r
    Eq. (9): KL target for the Bernoulli mask prior controls sufficiency mask sparsity; tuned per dataset.
  • loss weights λ_rec, λ_kld, λ_sp, λ_js, λ_mask, λ_con, λ_cos
    Stage I/II objectives in Eq. (8) and Sec. 6; numerical values are not reported and are chosen by hand.
  • maximum lag L and latent dimension d
    Sec. 5.1-5.3: lag window and latent dimensionality are assumed; values are not given.
  • intervention reference value z_tau,k
    Sec. 5.3: do(z_tau,k = z_tau,k) requires a reference value; not specified how it is chosen, and it affects all Δp scores.
  • evaluation time steps T_eval
    Sec. 6.2: necessity targets are computed only at T_eval; the choice is not stated.
assumptions (5)
  • domain assumption The observed time series is a nonlinear invertible mixture of latent factors with a temporal SCM containing instantaneous and lagged dependencies (Eq. 3).
    Sec. 5.1: fundamental modeling assumption; not tested.
  • domain assumption The latent SCM is identifiable from observational data under the conditions of Li et al. [2024], and the learned inverse mechanisms recover the true dependencies.
    Sec. 6.1 cites Li et al. for identifiability but does not state or verify the conditions; abstract calls the process 'identifiable'.
  • ad hoc to paper The first-order Taylor linearization of Eq. (5) is accurate for the interventions used (small perturbations, J_i_i^(0)(s) ≠ 0).
    Sec. 5.3 derives Eq. (6) without error bounds; no validation against exact counterfactual propagation.
  • domain assumption Intervening on latent factors and decoding produces counterfactual inputs that are in-distribution for the black-box model.
    Sec. 5.4 feeds decoded counterfactual samples to F; off-manifold samples would give meaningless Δp.
  • domain assumption The black-box classifier's decision depends on the same latent generative factors, so latent necessity transfers to input-space necessity.
    Sec. 5.4/6: otherwise latent Δp does not explain F's decision on observed X.
invented entities (1)
  • temporal latent factors z_t,i in a latent causal process
    purpose: serve as intervention targets for estimating counterfactual necessity of temporal factors
    Newly posited latent variables; no falsifiable handle outside the model and no identifiability proof.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Beyond Sufficiency: Time Series Explanation with Counterfactual Necessity." pith.science (2026). https://pith.science/paper/6NS35DHT

@misc{pith2026260721573,
  author       = {Pith},
  title        = {Pith review of: Beyond Sufficiency: Time Series Explanation with Counterfactual Necessity},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6NS35DHT}},
  note         = {Machine review of arXiv:2607.21573}
}
read the original abstract

Faithful explanations of time-series classifiers should identify subsequences that are not only sufficient to preserve a black-box model's prediction, but also necessary for maintaining it. However, existing sufficiency-oriented methods can assign high importance to spurious subsequences that support the prediction without being essential to the model's decision. We introduce \textbf{TimePNS}, a necessity-aware framework for time-series explanation. Inspired by Pearl's counterfactual notion of necessity, TimePNS assesses whether a temporal factor is necessary by intervening on it and measuring whether the original prediction is disrupted. The framework adopts a two-stage design. Stage I learns an identifiable causal generative process together with a sufficiency-oriented explanation mask. Stage II performs counterfactual interventions on temporal factors to derive necessity signals, which supervise a temporal gate that refines the initial explanation by suppressing non-essential components and emphasizing counterfactually necessary ones. Experiments on synthetic and real-world time-series benchmarks show that TimePNS more accurately identifies decision-critical subsequences and consistently improves sufficiency-necessity trade-offs over strong baselines.

Figures

Figures reproduced from arXiv: 2607.21573 by the authors.

Figure 1
Figure 1. Latent variable genera￾tive process. The black dashed lines represent the latent causal process, while the gray lines indicate the mixture process. We model a multivariate time series as a nonlinear observation of an underlying latent causal process, following [Li et al., 2024]. Let X = {xt} T t=1 denote an observed time series, where xt ∈ R D. Each observed variable is assumed to be generated from a latent state zt… view at source ↗
Figure 2
Figure 2. Real-world occlusion evaluation on Epilepsy, ERing, and WordRec. [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Ablation experiment. STimePNS \ Sbaseline; and positions selected by both methods, Sboth = Sbaseline ∩ STimePNS. We ablate each set by replacing its values with Gaussian samples estimated from the training data, and measure the resulting change in the ground-truth class probability, |∆pGT|. As shown in Figure 3a, TimePNS-exclusive positions induce larger prediction changes than baseline-exclusive positions in most s… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 4 linked inside Pith

  1. [1]

    arXiv preprint arXiv:2405.15325 , year=

    On the identification of temporally causal representation with instantaneous dependence , author=. arXiv preprint arXiv:2405.15325 , year=

    arXiv

  2. [2]

    Proceedings of the 41st International Conference on Machine Learning , pages=

    TIMEX++ learning time-series explanations with information bottleneck , author=. Proceedings of the 41st International Conference on Machine Learning , pages=

  3. [3]

    Advances in Neural Information Processing Systems , volume=

    Encoding time-series explanations through self-supervised model behavior consistency , author=. Advances in Neural Information Processing Systems , volume=

  4. [4]

    2009 , publisher=

    Causality , author=. 2009 , publisher=

    2009

  5. [5]

    Advances in neural information processing systems , volume=

    Unsupervised feature extraction by time-contrastive learning and nonlinear ica , author=. Advances in neural information processing systems , volume=

  6. [6]

    International Conference on Machine Learning , pages=

    Explaining time series predictions with dynamic masks , author=. International Conference on Machine Learning , pages=. 2021 , organization=

    2021

  7. [7]

    ACM Transactions on Management Information Systems , volume=

    Time series prediction using deep learning methods in healthcare , author=. ACM Transactions on Management Information Systems , volume=. 2023 , publisher=

    2023

  8. [8]

    Artificial Intelligence Review , volume=

    A comprehensive survey of deep learning for time series forecasting: architectural diversity and open challenges , author=. Artificial Intelligence Review , volume=. 2025 , publisher=

    2025

Show all 37 references
  1. [9]

    arXiv preprint arXiv:1611.01144 , year=

    Categorical reparameterization with gumbel-softmax , author=. arXiv preprint arXiv:1611.01144 , year=

    arXiv

  2. [10]

    International Conference on Machine Learning , pages=

    Towards trustworthy explanation: On causal rationalization , author=. International Conference on Machine Learning , pages=. 2023 , organization=

    2023

  3. [11]

    arXiv preprint arXiv:2506.09853 , year=

    Causal sufficiency and necessity improves chain-of-thought reasoning , author=. arXiv preprint arXiv:2506.09853 , year=

  4. [12]

    International Conference on Machine Learning , pages=

    Learning perturbations to explain time series predictions , author=. International Conference on Machine Learning , pages=. 2023 , organization=

    2023

  5. [13]

    2020 International joint conference on neural networks (IJCNN) , pages=

    Conceptual explanations of neural network prediction for time series , author=. 2020 International joint conference on neural networks (IJCNN) , pages=. 2020 , organization=

    2020

  6. [14]

    Advances in Neural Information Processing Systems , volume=

    Invariant learning via probability of sufficient and necessary causes , author=. Advances in Neural Information Processing Systems , volume=

  7. [15]

    Ieee Access , volume=

    Explainable AI for time series classification: a review, taxonomy and research directions , author=. Ieee Access , volume=. 2022 , publisher=

    2022

  8. [16]

    2025 IEEE International Conference on Data Mining (ICDM) , pages=

    Explanation space: A new perspective into time series interpretability , author=. 2025 IEEE International Conference on Data Mining (ICDM) , pages=. 2025 , organization=

    2025

  9. [17]

    Nature Machine Intelligence , volume=

    Evaluation of post-hoc interpretability methods in time-series classification , author=. Nature Machine Intelligence , volume=. 2023 , publisher=

    2023

  10. [18]

    The Thirteenth International Conference on Learning Representations , year=

    Shedding light on time series classification using interpretability gated networks , author=. The Thirteenth International Conference on Learning Representations , year=

  11. [19]

    Towards transparent time series forecasting , author=

  12. [20]

    Advances in Neural Information Processing Systems , volume=

    Abstracted shapes as tokens-a generalizable and interpretable model for time-series classification , author=. Advances in Neural Information Processing Systems , volume=

  13. [21]

    Joint european conference on machine learning and knowledge discovery in databases , pages=

    Ts-mule: Local interpretable model-agnostic explanations for time series forecast models , author=. Joint european conference on machine learning and knowledge discovery in databases , pages=. 2021 , organization=

    2021

  14. [22]

    International conference on machine learning , pages=

    Axiomatic attribution for deep networks , author=. International conference on machine learning , pages=. 2017 , organization=

    2017

  15. [23]

    arXiv preprint arXiv:2107.14317 , year=

    Temporal dependencies in feature importance for time series predictions , author=. arXiv preprint arXiv:2107.14317 , year=

    arXiv

  16. [24]

    Neural Networks , volume=

    Unifying invariant and variant features for graph out-of-distribution via probability of necessity and sufficiency , author=. Neural Networks , volume=. 2025 , publisher=

    2025

  17. [25]

    Proceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence , pages=

    Going beyond consistency: target-oriented multi-view graph neural network , author=. Proceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence , pages=

  18. [26]

    Proceedings of the IEEE international conference on computer vision , pages=

    Interpretable explanations of black boxes by meaningful perturbation , author=. Proceedings of the IEEE international conference on computer vision , pages=

  19. [27]

    arXiv preprint arXiv:2303.02794 , year=

    Cortx: Contrastive framework for real-time explanation , author=. arXiv preprint arXiv:2303.02794 , year=

    arXiv

  20. [28]

    Advances in Neural Information Processing Systems , volume=

    Improving deep learning interpretability by saliency guided training , author=. Advances in Neural Information Processing Systems , volume=

  21. [29]

    Advances in neural information processing systems , volume=

    Attention is all you need , author=. Advances in neural information processing systems , volume=

  22. [30]

    Journal of machine learning research , volume=

    Learning precise timing with LSTM recurrent networks , author=. Journal of machine learning research , volume=

  23. [31]

    arXiv preprint arXiv:1603.06995 , year=

    Multi-scale convolutional neural networks for time series classification , author=. arXiv preprint arXiv:1603.06995 , year=

    arXiv

  24. [32]

    IEEE engineering in medicine and biology magazine , volume=

    The impact of the MIT-BIH arrhythmia database , author=. IEEE engineering in medicine and biology magazine , volume=. 2001 , publisher=

    2001

  25. [33]

    IEEE/CAA Journal of Automatica Sinica , volume=

    The UCR time series archive , author=. IEEE/CAA Journal of Automatica Sinica , volume=. 2019 , publisher=

    2019

  26. [34]

    Proceedings of the IEEE , volume=

    Toward causal representation learning , author=. Proceedings of the IEEE , volume=. 2021 , publisher=

    2021

  27. [35]

    Artificial intelligence and statistics , pages=

    Nonlinear ICA of temporally dependent stationary sources , author=. Artificial intelligence and statistics , pages=. 2017 , organization=

    2017

  28. [36]

    Causality: Models, reasoning, and inference

    Judea Pearl. Causality: Models, reasoning, and inference. Cambridge: Cambridge University Press. 384 pp., 2000, ISBN 0521773628. , author=. Psychometrika , volume=. 2002 , publisher=

    2000

  29. [37]

    arXiv preprint arXiv:2510.20084 , year=

    ShapeX: Shapelet-Driven Post Hoc Explanations for Time Series Classification Models , author=. arXiv preprint arXiv:2510.20084 , year=

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.